Everything we trade is an argument about electrons

In 1865, a young British economist named William Stanley Jevons sat down to answer a practical policy question. At the time, Britain was the richest country on earth and its entire economy ran on coal. Steam engines had just gotten dramatically more efficient and the obvious prediction was that coal consumption would fall due to the technological upgrades.

Well it didn’t.

Jevons ran the numbers and couldn’t believe his findings. Britain was burning an order of magnitude more coal than a generation earlier, not less. Efficiency had made coal cheaper, cheaper coal had unlocked new uses, new uses had spawned new industries, and total demand had gone vertical.

He called this the paradox of fuel, but economists later called it the Jevons Paradox. It is one of the most reliably confirmed observations in the history of the profession, and roughly nobody in the AI conversation is quoting it.

But they should be. Just look at the MAG7 receipts:

• Microsoft paid to restart a nuclear reactor it had previously decommissioned.
• Amazon bought 960 megawatts of capacity next door to a 2.5-gigawatt nuclear station.
• Google signed for 500 megawatts of small modular reactors that don’t physically exist yet.
• Meta filed an RFP to purchase between one and four gigawatts of new nuclear capacity.

These are not energy companies. But somewhere in the last eighteen months, the shopping list quietly stopped being chips and started being electrons.

I’m going to argue one specific thing here, and it’s narrower than the maximalist version you may have heard. The one where everything in the economy is fundamentally energy, and any exchange is the abstraction you can track back to joules.

The smaller claim is this: In the next decade, the binding constraint on the frontier of the economy, and therefore the shadow price that every serious capital allocation decision is really being made against, is dispatchable electricity.

The wrappers we use to describe what we’re buying will look the same, but what they’re priced against will quietly change.

We’ll be using the joule for obvious reasons, and because it condenses nicely into a one-word version.

The water table

Picture the economy as a desert with a giant aquifer running under it. Every generation drills a new kind of well. The wells get deeper, the pumps get fancier, and the water that comes up gets marketed in progressively more abstract packaging. Sparkling brand, Artisanal provenance, oh or, Programmable tokenization.

It’s all the same water.

The last two centuries fit on one axis and the axis is watts.

• 1771: Cotton mills → Waterpower
• 1829: Railways → Coal
• 1875: Electrification → Coal + Hydro
• 1908: Ford’s assembly line → Crude oil
• 1971: Microprocessor → Grid-delivered electrons

Carlota Perez named the pattern and called them the five technological surges. Her framework is load-bearing for everything that follows in this essay. The shape is always the same: build new infrastructure on top of new primary energy, speculate wildly about what the infrastructure can do, watch the bubble pop, and then spend the next two or three decades actually putting the technology to work.

Each surge dressed its output up as something other than energy. Cotton calico. Passenger miles. Lightbulb-hours. Gasoline gallons. Megabits per second. What got sold was the convenience on top of the joules, not the joules.

That’s how you knew the well was working.

The wrappers got very good

Money was the first and best disguise.

Gold-backed currency was a physical claim on a metal that was hard to get out of the ground because getting it out of the ground took enormous amounts of energy. Fiat currency replaced the metal with a promise backed by the state’s ability to tax energy flows. Take the energy away and every dollar becomes a suggestion. The supply chain of sovereignty runs on watts, even when the textbooks pretend otherwise.

Credit was the better disguise, because credit has no physical referent at all. Compound interest is a mathematical function that assumes infinite growth inside a physical system that obeys the second law of thermodynamics.

Frederick Soddy pointed this out in 1926. He was a Nobel chemist, he had watched physicists split the atom, and he spent the back half of his career trying to warn economists that their equations implied an energy future the planet couldn’t deliver. The economists politely ignored him for a century. The arithmetic didn’t care.

Then came attention, the wrapper we built the entire twentieth century around. Advertising paid by eyeball-hours. Eyeball-hours supplied by humans who ate calories grown, processed, fertilized, harvested, shipped, and refrigerated on petroleum. Vaclav Smil has spent a career on this point. Modern agriculture, he writes, is a machine for turning fossil fuel into food, modern media is a machine for turning food into attention, and modern advertising is a machine for turning attention into revenue. Four steps from the well to the CPM.

Then the wrappers started getting faster.

Data is attention, captured, stored on spinning metal, and indexed with megawatts of cooling. Compute is data that moved, priced per floating-point operation and delivered per watt. Intelligence is the current wrapper. Priced per token today. We’ll come back to that one.

Notice the acceleration. Each new wrapper is shorter-lived than the last, cheaper to produce, and more obviously downstream of the generation capacity holding it up. We’re running out of abstractions faster than we can invent new ones.

Jevons keeps winning

Every efficiency gain since 1865 has done the same thing Jevons saw in the coal data.

Electric lighting was supposed to reduce the energy we spent on lighting. Global lighting energy consumption went up. Jet engines got twice as efficient per passenger-mile. Global aviation fuel use quadrupled. Air conditioning got dramatically more efficient per BTU removed. Cooling is now the single fastest-growing residential electricity load on earth. Data centers got better at compute per watt by roughly three orders of magnitude between 2000 and 2020. Total data center power consumption went from statistical noise to ten percent of US electricity demand.

The story we keep telling ourselves, that abstraction and efficiency let us escape the constraint, has always been wrong in exactly the same way. The constraint never went anywhere. We just found clever new ways to buy more of what it was rationing.

So when someone tells you AI will decouple value from energy because inference keeps getting cheaper, you should ask how the previous nineteen times that story got told actually worked out.

Landauer, briefly

There is a floor under all of this, and Rolf Landauer worked it out in 1961.

Erasing a single bit of information at room temperature costs at least 2.87 × 10⁻²¹ joules. That’s the thermodynamic minimum for forgetting something. It is staggeringly small, and it is also a law, not a target. Modern GPUs sit about ten orders of magnitude above that floor. A ten-billion-times gap. Which is either terrifying or thrilling depending on whether you’re paying the power bill or writing the inference-cost roadmap. Compute can get a trillion times cheaper, in joules per bit, before physics calls time.

That’s the floor. The wrappers we’ve built are everything between the floor and the ceiling. The ceiling is the grid.

The bill comes due

This is where the historical argument collides with the present.

For the first time since Perez named the pattern, a surge is hitting its energy ceiling before its efficiency ceiling. The IEA’s base-case projection has global data-center electricity going from 415 terawatt-hours in 2024 to 945 terawatt-hours by 2030. That’s roughly Japan’s entire annual electricity consumption, appearing on the grid, to run matrix multiplications. Lawrence Berkeley’s 2024 report puts US data centers at 4.4 percent of national electricity in 2023 and projects them to land between roughly seven and twelve percent by 2028. Epoch AI’s team has the stock of high-quality text on the open web running out somewhere between 2026 and 2032 at current training rates.

Then come the deals. Microsoft restarted Three Mile Island Unit 1 for 835 megawatts on a twenty-year PPA. Amazon signed 960 megawatts at Talen’s Cumulus site. Google committed to 500 megawatts of Kairos SMRs through 2035. Meta’s RFP is shopping for one to four gigawatts of nuclear. xAI’s Memphis cluster hit interconnection limits within months and is running gas turbines on-site while it waits for the grid to catch up.

The interconnection queue is the quiet giveaway. In most US ISO regions, a new large generation project is now roughly five years out from application to first power, up from under two in 2008. Every hyperscaler with public ambitions for 2028 was already in that queue in 2023. They priced the joules first and the silicon second.

AI is the last wrapper, not the point

AI gets top billing in this essay because it’s the wrapper currently in the process of tearing. But the thesis is not about AI. The thesis is that every wrapper we’ve ever built has eventually come off.

If text is running out, the marginal training datum comes from sensors in physical space. A Tesla FSD mile, a Waymo driverless mile, an hour of a Figure humanoid moving crates in a BMW factory. Adam Jonas, who runs Morgan Stanley’s global autos and shared mobility research, has spent two years arguing this is the whole next act. He calls it Physical AI, and his team’s “Humanoid 100” report maps the supply chain. Nvidia’s Jensen Huang made the same argument from the CES and GTC keynote stages in 2025, telling the room that “the ChatGPT moment for general robotics is just around the corner.” Every one of those data-generating endpoints is a sensor attached to an actuator attached to a battery attached to a grid. Every bit of embodied training data is a derivative of a joule.

So the wrapper peels one more layer, and what’s underneath is the same thing that was underneath the cotton mill in 1771, the locomotive in 1829, and the incandescent bulb in 1879. The aquifer hasn’t moved.

Where this argument has limits

I want to be honest about where this argument has limits, because I’ve seen too many “X is just Y” takes that grab the scariest stat and ignore everything pushing back on them. There are seven serious objections. All seven are real.

One, exergy is not energy. A joule of ambient air at room temperature is economically useless. The scarce thing is useful work, what physicists call exergy, and in our context that means dispatchable low-entropy power. A sloppier version of this thesis collapses the moment you remember that not all joules are equivalent.

Two, Shannon is not Landauer. The value of a piece of information has almost nothing to do with the cost of generating it. A private key is cheap to compute and very expensive to find. Information economics and thermodynamic economics run on different tracks.

Three, coordination goods. Trust, law, brand, status, network effects. None of these obviously price in joules, and plausibly none of them ever will. A trademark has almost no relationship to the kilowatt-hours burned to establish it.

Four, land and time. Ricardian land rent and human time preference are irreducible scarcities. A beachfront house and a calendar year are valuable for reasons that have nothing to do with energy.

Five, the dematerialization numbers. Real US GDP per unit of energy has roughly doubled since 1990, per EIA. If you squint, the decoupling argues against this entire essay. My honest read is that most of the decoupling is accounting, because we offshored the energy-intensive stuff to China, and Smil has the receipts. But I can’t prove that in a blog post.

Six, Nicholas Georgescu-Roegen, the intellectual grandfather of this argument, explicitly warned against exactly the kind of monistic reduction the thesis flirts with. He’s my best ally and he’d push back. Good allies do that.

Seven, the Physical AI transition might take twenty years instead of five. Jonas is speculating about a pace of adoption. Humanoids might land as narrow industrial tools, not general-purpose agents. If the physical pivot is slower or weirder than the current consensus, a lot of the capital-allocation urgency in this argument softens. The grid constraint is still real. The shadow price is still there. But the 2028 timeline people are underwriting against might turn into a 2038 timeline, and that changes what you do about it tomorrow.

Take all seven seriously. After you do, one claim survives, and it’s the narrower one.

The binding constraint and shadow price behind serious capital allocation is already becoming dispatchable power, and the wrappers we use on top will increasingly track that shadow price whether they say so or not.

That’s the claim, and it’s the small one. The next trillion-dollar category is being priced against electrons at the margin, whatever the invoice says.

Landing the plane

Here’s the operator version, because this is a blog for operators.

Perez’s framework says every surge has two halves. Installation, driven by financial capital and bubbles, and deployment, driven by production capital and buildout. The two halves are separated by a crisis. We are in the deployment phase of the ICT surge, with AI as its final act. Perez herself has been writing for years about the overlap between this deployment phase and the clean-energy transition, which is why her recent work reads less like economic history and more like a shopping list for 2030. The carrier of this deployment phase is the electron. The nuclear PPAs are the canary.

Jonas’s Physical AI thesis is the second half of the same story, viewed from the P&L side. When AI leaves the data center, the economics of intelligence collapse into two line items. Cost of electricity. Cost of capital. His TAM numbers for humanoids are speculative, and he’d tell you that himself. The direction is not.

If intelligence becomes commodity, and it will, the moat stops being the model and becomes the megawatt. Every company that sells intelligence at the margin is going to look, five years from now, like an energy arbitrageur with a software front end. The nuclear deals are the early form. The humanoid pilots are the second form. The third form we haven’t named yet.

It’s all electrons

The wells are getting deeper. From down there you can see the water.

Two hundred years of wrappers. Cotton, steam, electricity, oil, silicon, tokens. Every one of them sold as something other than what it was. Every one of them priced against something that looked more and more abstract until the abstraction cracked and the aquifer came back into view. Jevons saw the pattern in coal in 1865. Soddy carried it forward to credit. Smil has spent forty years watching the same logic eat food, light, and compute.

Now it’s the AI wrapper’s turn. The disguise is good. The water underneath hasn’t moved.

It’s all electrons.